AUTOMATED NEAR REAL TIME RADARSAT-2 IMAGE GEO-PROCESSING
AND ITS APPLICATION FOR SEA ICE AND OIL SPILL MONITORING
Ziqiang Ou
Canadian Ice Service, Environment Canada
373 Sussex Drive, Ottawa, Canada K1A 0H3 - Ziqiang.Ou@ec.gc.ca
KEY WORDS: RADARSAT, SAR, Georeferencing, Real-time, Pollution, Snow Ice, Detection
ABSTRACT:
Synthetic aperture radar (SAR) data from RADARSAT-1 are an important operational data source for several ice centres around the
world. Sea ice monitoring has been the most successful real-time operational application for RADARSAT-1 data. These data are
used by several national ice services as a mainstay of their programs. Most recently, Canadian Ice Service launched another new
real-time operational program called ISTOP (Integrated Satellite Tracking Of Pollution). It uses the RADARSAT-1 data to monitor
ocean and lakes for oil slicks and tracking the polluters. In this paper, we consider the ice information requirements for operational
sea ice monitoring and the oil slick and target detection requirements for operational ISTOP program at the Canadian Ice Service
(CIS), and the potential for RADARSAT-2 to meet those requirements. Primary parameters are ice-edge location, ice concentration,
and stage of development; secondary parameters include leads, ice thickness, ice topography and roughness, ice decay, and snow
properties. Iceberg detection is an additional ice information requirement. The oil slick and ship target detection are the basic
requirements for the surveillance and pollution control.
1. INTRODUCTION
2. GEOMETRIC TRANSFORMATION
RADARSAT-2 products are provided in a GeoTIFF format
with a sets of support XML files for all products except the
RAW. The GeoTIFF images are georeferenced, but not
geocorrected so that they are not ready for most of GIS and
Remote Sensing applications, which require the RADARSAT-2
images being overlaid with other geographic data layers. Since
the products are not geographically corrected, the geographic
metadata included in GeoTIFF is limited to a set of points tying
image location to geographic location. GeoTIFF images is
generated in TIFF strip format. Multipolar images is generated
as separate GeoTIFF image files. All images are oriented such
that north is nominally up and east is nominally on the right.
Further processing is required in order to integrate the dataset
into CIS existing operational Integrated Spatial Information
System – ISIS (Ou, 2004; Koonar et al. 2004) and the
Integrated Satellite Tracking Of Pollution – ISTOP system
(Gauthier et al. 2007).
There are two major techniques for correcting geometric
distortion present in an image. The first is to model the nature
and magnitude of the sources of distortion and then use these
models to correct for distortions. The second is to establish a
mathematical relationship between the locations of pixels in an
image (X,Y) and their location (U,V) in the real world (e.g. on a
map). The latter is by far the most common, though many
standard sources of satellite images will have at least some
corrections done by modelling before the user receives the data.
Both ISIS for ice monitoring and ISTOP for oil monitoring are
real-time or near real-time applications. The data acquisition
and the data processing time is one of the critical factors.
Considering the data volume of SAR images and in order to
meet the time requirements, an automated and fast SAR image
processing and geo-referencing algorithm is the key for the
success of both applications. Thus, this paper will discuss the
automated fast RADARSAT-2 image geo-processing by using
the image geographical tie points provided in the RADARSAT2 product; the evaluation of geographical location accuracy for
the geocoded image; the evaluation of image geo-spatial
distortion introduced by geo-processing and how to minimize
the distortion and improve the image’s geo-location accuracy.
A set of different geo-reference procedures will be evaluated
based on their spatial accuracy of geocoded image and their
computational efficiencies in order to meet the requirements for
the mapping accuracy of ice and oil and the time constraint of
these real-time applications.
The idea is to develop a mathematical function that relates the
image and real world coordinate systems. e.g. X = f(U, V) and
Y = g(U, V). While in the end we want to transform our images
in the pixel coordinate system into some coordinate system in
the real world, in actuality this is accomplished in reverse. Once
a real world coordinate system is established, the mapping
functions (f, g) are used to determine which pixels in the image
whose centers fall closest to the locations of the real-world grid.
The transformation from a image coordinate system to a real
world coordinate system through a set of tie points of the two
coordinate systems can be mathematically expressed by a set of
mapping functions with respect to two coordinate components.
The coordinate mapping from any point [u, v] in the image
coordinate system to the corresponding point [x, y] in the real
world coordinate system is modelled as:
x = f (u, v / α )
y = g (u, v / β )
(1)
where α and β are parameters of the mapping function f and g
respectively and are determined by control points, and are
commonly determined by a global transformation.
1073
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
⎡α 0 ⎤
⎡β 0 ⎤
⎢
⎢β ⎥
⎥
⎡ε x 1 ⎤
⎡ε y 1 ⎤
α1 ⎥
⎢
⎢ 1⎥
⎢ε ⎥
⎢ε ⎥
⎢
⎢β ⎥
⎥
α
x
y
εx = ⎢ 2 ⎥ , εy = ⎢ 2 ⎥ , A = ⎢ 2⎥ , B = ⎢ 2⎥
⎢ # ⎥
⎢ # ⎥
⎢α 3 ⎥
⎢β 3 ⎥
⎢ ⎥
⎢ ⎥
⎢α ⎥
⎢β ⎥
⎣⎢ε xn ⎦⎥
⎣⎢ε yn ⎥⎦
⎢ 4⎥
⎢ 4⎥
⎢⎣α 5 ⎥⎦
⎢⎣ β 5 ⎥⎦
Since the 2D polynomial functions do not reflect the sources of
distortion during the image formation and do not correct for
terrain relief distortions, they are limited to images with few or
small distortions, such as nadir-viewing images, systematicallycorrected images and/or small images over flat terrain. These
functions correct for local distortions at the ground control point
(GCP) location. They are very sensitive to input errors and
hence GCPs have to be numerous and regularly distributed.
Consequently, these functions should not be used when precise
geometric positioning is required for multi-source multi-format
data integration and in high relief areas. If a second order
polynomial function is used, the following is obtained:
x = α 0 + α 1u + α 2 v + α 3uv + α 4 u 2 + α 5 v 2
⎡1 u1
⎢
1 u2
W =⎢
⎢# #
⎢
⎣⎢1 u n
(2)
y = β 0 + β 1u + β 2 v + β 3 uv + β 4 u 2 + β 5 v 2
From a mathematical point of view, a first order polynomial
transformation requires a minimum of 3 points, a second order
polynomial requires a minimum of 6 points and, a third order
polynomial requires a minimum of 10 points. Generally, if the
order of polynomial model is n, we must at least have a set of
M=(n+1)(n+2)/2 GCPs to solve Equations (1). Suppose that {(ui,
vi): i = 1, …, n} and {(xi, yi): i=1, …, n} are the tying GCP
pairs from image and real world coordinate systems
respectively, we can form n equations for both x and y as
following:
2
v1
v2
#
u1v1
u 2 v2
#
u1
2
u2
#
vn
u n vn
un
2
2
v1 ⎤
2⎥
v2 ⎥
# ⎥
2⎥
vn ⎦⎥
⎡ x1 ⎤
⎡ y1 ⎤
⎢x ⎥
⎢y ⎥
2⎥
2
⎢
X=
, Y =⎢ ⎥
⎢#⎥
⎢#⎥
⎢ ⎥
⎢ ⎥
⎣ xn ⎦
⎣ yn ⎦
This means that the two sets of coefficients αj and βj could be
estimated separately. The least square estimation can be used to
determine the coefficients of αj and βj from GCPs.
x i = α 0 + α 1u i + α 2 v i + α 3 u i v i + α 4 u i + α 5 v i + ε x i
2
2
2
2
y i = β 0 + β1u i + β 2 vi + β 3 u i vi + β 4 u i + β 5 vi + ε y i Min[ε xT ε x ] = (WAˆ − X ) T (WAˆ − X )
(3)
Min[ε T ε ] = (WBˆ − Y ) T (WBˆ − Y )
y
By assuming that the ε are independent and identically
distributed (iid) observation errors, i.e.,
⎧⎪⎡0⎤ ⎡σ 2
⎡ε x ⎤
0 ⎤ ⎫⎪
ε = ⎢ ⎥ , ε ~ N ⎨⎢ ⎥ , ⎢ x
2 ⎥⎬
⎪⎩⎣0⎦ ⎢⎣ 0 σ y ⎥⎦ ⎪⎭
⎣ε y ⎦
X = WA + ε x
Y = WB + ε y
y
These are two standard least square problems. The estimation
for αj and βj are in Equation (7).
Aˆ = (W T W ) −1W T X
Bˆ = (W T W ) −1W T Y
(4)
the equations constructed from the GCPs by (3) could be
separated into two sets of equations as shown in (5) below,
(6)
(7)
The corresponding variance components estimation (Ou, 1989)
are in Equation (8).
(5)
σˆ x 2 =
1
n−r
σˆ y 2 =
1
n−r
(WAˆ − X ) T (WAˆ − X )
(WBˆ − Y ) T (WBˆ − Y )
(8)
where
where r = rank(W). Thus, the transformation functions f and g
are determined. The total variance could be estimated by:
σˆ t2 = σˆ x2 + σˆ y2
1074
(9)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Sometimes we also need to determine the inverse
transformation of Equation (1) from the GCPs set, which
transforms the coordinate from (xi, yi) to (ui, vi).
u = f −1 ( x, y / α ′)
(10)
v = g −1 ( x, y / β ′)
The same procedure could be used to estimate the coefficients
α’j and β’j of transformation (10).
There are many challenges presented by global transformations
that require consideration here. For global transformation
models such as affine transformation, projective transformation,
and global polynomial transformation, the parameters α and β
are the same for all the points and are determined by all the
control points, so a single function is used to model the
transformation for each component of coordinates. When the
geometric distortion is complex and location dependent, global
models become inadequate to model the image geometry.
Linear functions such as affine transformation and projective
transformation are too simple to take the local variation into
consideration. Nonlinear functions such as global polynomials
use the least square methods to optimize the parameters, thus
the local variation will be averaged across the whole image
(Zitova and Flusser, 2003). Consequently, the registration error
of locally deformed images by global functions is usually large
and the spatial distribution of the error also varies with the
location.
All products (except RAW) are georeferenced, but not
geocorrected. Since these products are not geographically
corrected, the geographic metadata included in GeoTIFF is
limited to the four corners tying image location to geographic
location. GeoTIFF images are generated in TIFF strip format.
Multi-polar images will be generated as separate GeoTIFF
image files. All images are oriented such that north is nominally
up and east is nominally on the right.
A grid of tie points is included in the product.xml under the
geolocationGrid node (highlighted in Figure 2), which ties the
line/pixel positions in image coordinates to geographical
latitude/longitude. The image coordinates are in units of pixels.
The ground coordinates are latitude and longitude in units of
decimal degrees. The ground coordinates are referenced to
WGS-84, and pixel and line numbers start at 0. These grid tie
points are used as Ground Control Points (GCP) to
automatically geo-correct the image.
3. RADARSAT-2 IMAGE GEOPROCESSING
The basic product as generated by the RADARSAT-2 processor
contains a Product Information File and one or more Image
Pixel Data Files. The composition of RADARSAT-2 products
is shown in Figure 1. All RADARSAT-2 products include one
or more Image Pixel Data Files. One, two, or four Image Pixel
Data Files may be included, corresponding to single, dual, or
quad polarization modes, respectively. Each file contains the
raster SAR image for a given polarization in GeoTIFF format
(MDA, 2003).
The Figure 3 and Figure 4 below give an example of
RADARSAT-2 ScanSAR Wide dataset with image size of
10508 × 10039 and pixel spacing and line spacing of 50 meters.
The Figure 3 shows the GCP distribution in image coordinates
and Figure 4 shows the GCP distribution in geographic
coordinates. The GCPs are evenly distributed across the whole
image.
Product
1
Product
Information
File
1..4
Image
Pixel
Data File
Figure 2: RADARSAT-2 Product XML Node Tree
1..n
Support
Files
Figure 1: Product Composition
The Product Information File is an ASCII file that logically
groups known information on the product. For example,
groupings are provided for source, image generation and
imagery information related to the product. The Product
Information File is encoded in Extensible Markup Language
(XML) format as shown in Figure 2.
1075
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Figure 3: Sample of RADARSAT-2 GCPs Distribution in
Image Coordinate System
Figure 5:
First, the geographic latitude and longitude of GCPs are
projected onto a planar map coordinate system. In our case, we
use Polar Stereographic (PS) projection for high arctic region
and Lambert Conformal Conic (LCC) for the rest of Canada.
Then, a global polynomial transformation Equation (2) between
the image coordinates (pixel/line) and the projected map
coordinates is determined through the least square regression.
In most cast, a perfect fit for all GCPs would require an
unnecessarily high order of transformation. Instead of
increasing the order, the user has option to tolerate a certain
amount of error. When the transformation coefficients of the
global polynomial transformation equation (1) and its the
inverse transformation (10) are calculated, the inverse
transformation (10) is used to retransform the reference
coordinates of the GCPs back to the source coordinate system.
Unless the order of transformation allows for a perfect fit, there
is some discrepancy between the source coordinates and the
retransformed reference coordinates. The Figure 5 displays the
discrepancies at all GCP locations for the example in Figure 3.
It clearly shows that the errors along the image edges are larger
than the centre of the image.
Total Errors for a 2nd Order Global Polynomial
Transformation (RMS: 0.96 Pixel)
The Root Mean Square (RMS) error is the average distance
between the input source location of all GCPs and their
retransformed locations. They can be estimated based on
equation (8) and (9). The discrepancies shown in Figure 5 are
enlarged for presentation purpose. The actual values are
estimated as below.
σˆ x = 0.66 , σˆ y = 0.70
Figure 4: GCPs Geographic Distribution
and
σˆ t = 0.96
(11)
The estimated total RMS error is 0.96 pixel, which is smaller
than 1 pixel. This indicates that a second order global
transformation polynomial is good enough for our ice and oil
spill monitoring applications. The geolocationGrid tie points
(GCPs) provided in product.xml will be used to automatically
geo-correct the RADARSAT-2 images. The diagram in Figure
6 shows the automated data processing flow.
1076
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Incoming
Files
RSAT-2 Data
Geoprocessing
Receiving
Station
Database
ISIS
Ice Forecast
Workstation
Products
Dissemination
ISTOP
Oil Spill
Workstation
Co-Polarization Channel: VV
Cross-Polarization Channel: VH
Figure 7: Co-Polarization vs Cross-Polarization at High Wind
Condition
Figure 6: Data Processing Flow Diagram
5. OIL SPILL APPLICATIONS
4. ICE APPLICATIONS
Comparing to RADARSAT-1 single polarization SAR data,
RADARSAT-2 is capable of acquiring dual-polarization and
fully polarimetric SAR data. The RADARSAT-2 dualpolarization options include HH–HV or VV–VH modes. The
additional information provided by the cross-polarization
channel could be very useful, as the cross polarization channel
responds to different scattering mechanisms than the copolarization channel (Scheuchl et al. 2004). These dualpolarization data are available in all beam configurations,
giving a wide choice in resolution, coverage, and incidence
angle. Most importantly, dual polarization is available in
ScanSAR modes, which are mostly used for ice and oil
monitoring at CIS from RADARSAT-1 data. Operational
experience by CIS analysts results in a preference of HH over
VV for co-polarization channels.
The cross-polarization HV and VH backscatter response from
water are generally low and are relatively independent of windinduced surface roughness conditions, whereas their backscatter
from sea ice are affected by surface roughness, volume
scattering, and multi-bounce scattering. Thus, the ice–ocean
contrast of cross-polarization can be expected to be greater than
that for either of the co-polarization channels, especially at high
wind conditions as shown in Figure 7. The cross-polarization
data can enhance the structural information of sea ice and have
demonstrated some utility for improving discrimination
between smooth and deformed ice as well. The combined use of
co-polarization and cross-polarization channel will give better
results across a wider range of incidence angles.
CIS is an important partner to Canada’s NASP (National Aerial
Surveillance Program) by using earth observation technology
(RADARSAT imagery) to look for oil-like signatures
(anomalies) on the ocean’s surface that could be indicative of
an oil spill. The CIS operational ISTOP (Integrated Satellite
Tracking of Pollution) program currently uses RADARSAT-1
ScanSAR HH polarization data to identify potential oil spills, to
track ship targets (Gauthier et al. 2007), and to direct pollution
surveillance flights to the locations of potential pollution
incidents.
The SAR signal is sensitive to the roughness of the sea surface,
which is modulated by wind speed and direction; imagery
acquired at VV polarization is the most sensitive to wind speed
variability (as shown in Figure 7). The suppression of the
capillary waves by oil from either anthropogenic sources, such
as an oil spill, or from natural biological slicks, reduces the
surface roughness resulting in less radar backscatter and darker
image tones. The detection of oil slicks has been found to be
best in moderate wind conditions in the range of 3 to 10 m/s.
Although imagery in both VV and HH polarizations from
RADARSAT-2 can be used for slick detection, the VV imagery
might be preferred as, in general, it offers a better signal to
clutter ratio than other polarization choices (i.e., HH, VH, or
HV). Although VV is more sensitive than HH for slick
detection, there may not be any advantage to using the copolarized or cross-polarized signatures as oil-free and oilcovered surfaces tend to have similar contrast and polarization
ratios. Slick thickness and the inability to differentiate oil slicks
from "look-alikes" such as areas of low-wind, grease ice, or
biological surfactants remain problematic.
One of CIS research priority is to assess the utility of
RADARSAT-2 VV/VH dual channel ScanSAR data for the
ISTOP program. It is known that VV polarization provides
superior CNR for oil detection over HH polarization. On the
other hand, VV is less appropriate for ship detection. However,
VH polarization has shown promise in detecting ships (figure
3)[2]. ISTOP will be working with others to investigate more
fully whether the VH channel of RADARSAT-2 can perform
1077
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
adequately for ship detection. If so, VV/VH will be likely
replace HH as the default ISTOP acquisition mode (DeAbreu et
al. 2006).
Making, 2004 Swets & Zeitlinger, Lisse, ISBN:9058096521,
pp.221-227
MDA, 2003. RADARSAT-2 Product Format Definition, RNRP-51-2713 Issue 1/5: NOV. 24, 2003
REFERENCES
DeAbreu, R., Gauthier, M.F., and Wychen, W.V., 2006. SARBased Oil Pollution Surveillance in Canada: Operational
Implementation and Research Priorities, Proceedings
OceanSAR 2006 – Third Workshop on Coastal and Marine
Applications of SAR, St. John’s, NL, Canada, October 2006.
Gauthier, M. F., Weir, L., Ou, Z., Arkett, M. and DeAbreu, R.,
2007. Integrated Satellite Tracking of Pollution – A New
Operational Program, International Geoscience and Remote
Sensing Symposium (IGARSS), July 2007, Barcelona, Spain.
Koonar, A., Ou, Z., and Scarlett, B. 2004. A real-time geospatial information system: an integration of GIS technologies
for ice forecasting, Advances in Spatial Analysis and Decision
Ou, Z., 1989. Estimation of variance and covariance
components, Bulletin Geodesique, Vol.63, p.139-148.
Ou, Z., 2004. An Integrated Spatial Information System for Ice
Service, The 20th Congress of International Society of
Photogrammetry and Remote Sensing (ISPRS) at Istanbul,
Turkey.
Scheuchl, B., Flett, D., Ron Caves, R. and Cumming, I., 2004.
Potential of RADARSAT-2 data for operational sea ice
monitoring, Canadian Journal of Remote Sensing, Vol. 30, No.
3, pp. 448–461
Zitova, B. and Flusser, J., 2003, Image registration methods: a
survey, Image and Vision Computing, 21(11): 977-1000.
1078